src/gallium/drivers/freedreno/freedreno_batch.h

   1 /*
   2  * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #ifndef FREEDRENO_BATCH_H_
  28 #define FREEDRENO_BATCH_H_
  29
  30 #include "util/u_inlines.h"
  31 #include "util/u_queue.h"
  32 #include "util/list.h"
  33
  34 #include "freedreno_util.h"
  35
  36 struct fd_context;
  37 struct fd_resource;
  38 enum fd_resource_status;
  39
  40 /* Bitmask of stages in rendering that a particular query query is
  41  * active.  Queries will be automatically started/stopped (generating
  42  * additional fd_hw_sample_period's) on entrance/exit from stages that
  43  * are applicable to the query.
  44  *
  45  * NOTE: set the stage to NULL at end of IB to ensure no query is still
  46  * active.  Things aren't going to work out the way you want if a query
  47  * is active across IB's (or between tile IB and draw IB)
  48  */
  49 enum fd_render_stage {
  50         FD_STAGE_NULL     = 0x01,
  51         FD_STAGE_DRAW     = 0x02,
  52         FD_STAGE_CLEAR    = 0x04,
  53         /* used for driver internal draws (ie. util_blitter_blit()): */
  54         FD_STAGE_BLIT     = 0x08,
  55         FD_STAGE_ALL      = 0xff,
  56 };
  57
  58 #define MAX_HW_SAMPLE_PROVIDERS 5
  59 struct fd_hw_sample_provider;
  60 struct fd_hw_sample;
  61
  62 /* A batch tracks everything about a cmdstream batch/submit, including the
  63  * ringbuffers used for binning, draw, and gmem cmds, list of associated
  64  * fd_resource-s, etc.
  65  */
  66 struct fd_batch {
  67         struct pipe_reference reference;
  68         unsigned seqno;
  69         unsigned idx;       /* index into cache->batches[] */
  70
  71         int in_fence_fd;
  72         bool needs_out_fence_fd;
  73         struct pipe_fence_handle *fence;
  74
  75         struct fd_context *ctx;
  76
  77         struct util_queue_fence flush_fence;
  78
  79         /* do we need to mem2gmem before rendering.  We don't, if for example,
  80          * there was a glClear() that invalidated the entire previous buffer
  81          * contents.  Keep track of which buffer(s) are cleared, or needs
  82          * restore.  Masks of PIPE_CLEAR_*
  83          *
  84          * The 'cleared' bits will be set for buffers which are *entirely*
  85          * cleared, and 'partial_cleared' bits will be set if you must
  86          * check cleared_scissor.
  87          */
  88         enum {
  89                 /* align bitmask values w/ PIPE_CLEAR_*.. since that is convenient.. */
  90                 FD_BUFFER_COLOR   = PIPE_CLEAR_COLOR,
  91                 FD_BUFFER_DEPTH   = PIPE_CLEAR_DEPTH,
  92                 FD_BUFFER_STENCIL = PIPE_CLEAR_STENCIL,
  93                 FD_BUFFER_ALL     = FD_BUFFER_COLOR | FD_BUFFER_DEPTH | FD_BUFFER_STENCIL,
  94         } cleared, partial_cleared, restore, resolve;
  95
  96         /* is this a non-draw batch (ie compute/blit which has no pfb state)? */
  97         bool nondraw : 1;
  98         bool needs_flush : 1;
  99         bool blit : 1;
 100         bool back_blit : 1;      /* only blit so far is resource shadowing back-blit */
 101
 102         /* Keep track if WAIT_FOR_IDLE is needed for registers we need
 103          * to update via RMW:
 104          */
 105         bool needs_wfi : 1;
 106
 107         /* To decide whether to render to system memory, keep track of the
 108          * number of draws, and whether any of them require multisample,
 109          * depth_test (or depth write), stencil_test, blending, and
 110          * color_logic_Op (since those functions are disabled when by-
 111          * passing GMEM.
 112          */
 113         enum {
 114                 FD_GMEM_CLEARS_DEPTH_STENCIL = 0x01,
 115                 FD_GMEM_DEPTH_ENABLED        = 0x02,
 116                 FD_GMEM_STENCIL_ENABLED      = 0x04,
 117
 118                 FD_GMEM_MSAA_ENABLED         = 0x08,
 119                 FD_GMEM_BLEND_ENABLED        = 0x10,
 120                 FD_GMEM_LOGICOP_ENABLED      = 0x20,
 121         } gmem_reason;
 122         unsigned num_draws;   /* number of draws in current batch */
 123
 124         /* Track the maximal bounds of the scissor of all the draws within a
 125          * batch.  Used at the tile rendering step (fd_gmem_render_tiles(),
 126          * mem2gmem/gmem2mem) to avoid needlessly moving data in/out of gmem.
 127          */
 128         struct pipe_scissor_state max_scissor;
 129
 130         /* Track the cleared scissor for color/depth/stencil, so we know
 131          * which, if any, tiles need to be restored (mem2gmem).  Only valid
 132          * if the corresponding bit in ctx->cleared is set.
 133          */
 134         struct {
 135                 struct pipe_scissor_state color, depth, stencil;
 136         } cleared_scissor;
 137
 138         /* Keep track of DRAW initiators that need to be patched up depending
 139          * on whether we using binning or not:
 140          */
 141         struct util_dynarray draw_patches;
 142
 143         /* Keep track of writes to RB_RENDER_CONTROL which need to be patched
 144          * once we know whether or not to use GMEM, and GMEM tile pitch.
 145          *
 146          * (only for a3xx.. but having gen specific subclasses of fd_batch
 147          * seemed overkill for now)
 148          */
 149         struct util_dynarray rbrc_patches;
 150
 151         struct pipe_framebuffer_state framebuffer;
 152
 153         /** draw pass cmdstream: */
 154         struct fd_ringbuffer *draw;
 155         /** binning pass cmdstream: */
 156         struct fd_ringbuffer *binning;
 157         /** tiling/gmem (IB0) cmdstream: */
 158         struct fd_ringbuffer *gmem;
 159
 160         // TODO maybe more generically split out clear and clear_binning rings?
 161         struct fd_ringbuffer *lrz_clear;
 162
 163         /**
 164          * hw query related state:
 165          */
 166         /*@{*/
 167         /* next sample offset.. incremented for each sample in the batch/
 168          * submit, reset to zero on next submit.
 169          */
 170         uint32_t next_sample_offset;
 171
 172         /* cached samples (in case multiple queries need to reference
 173          * the same sample snapshot)
 174          */
 175         struct fd_hw_sample *sample_cache[MAX_HW_SAMPLE_PROVIDERS];
 176
 177         /* which sample providers were active in the current batch: */
 178         uint32_t active_providers;
 179
 180         /* tracking for current stage, to know when to start/stop
 181          * any active queries:
 182          */
 183         enum fd_render_stage stage;
 184
 185         /* list of samples in current batch: */
 186         struct util_dynarray samples;
 187
 188         /* current query result bo and tile stride: */
 189         struct pipe_resource *query_buf;
 190         uint32_t query_tile_stride;
 191         /*@}*/
 192
 193
 194         /* Set of resources used by currently-unsubmitted batch (read or
 195          * write).. does not hold a reference to the resource.
 196          */
 197         struct set *resources;
 198
 199         /** key in batch-cache (if not null): */
 200         const void *key;
 201         uint32_t hash;
 202
 203         /** set of dependent batches.. holds refs to dependent batches: */
 204         uint32_t dependents_mask;
 205 };
 206
 207 struct fd_batch * fd_batch_create(struct fd_context *ctx, bool nondraw);
 208
 209 void fd_batch_reset(struct fd_batch *batch);
 210 void fd_batch_sync(struct fd_batch *batch);
 211 void fd_batch_flush(struct fd_batch *batch, bool sync, bool force);
 212 void fd_batch_add_dep(struct fd_batch *batch, struct fd_batch *dep);
 213 void fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc, bool write);
 214 void fd_batch_check_size(struct fd_batch *batch);
 215
 216 /* not called directly: */
 217 void __fd_batch_describe(char* buf, const struct fd_batch *batch);
 218 void __fd_batch_destroy(struct fd_batch *batch);
 219
 220 /*
 221  * NOTE the rule is, you need to hold the screen->lock when destroying
 222  * a batch..  so either use fd_batch_reference() (which grabs the lock
 223  * for you) if you don't hold the lock, or fd_batch_reference_locked()
 224  * if you do hold the lock.
 225  *
 226  * WARNING the _locked() version can briefly drop the lock.  Without
 227  * recursive mutexes, I'm not sure there is much else we can do (since
 228  * __fd_batch_destroy() needs to unref resources)
 229  */
 230
 231 static inline void
 232 fd_batch_reference(struct fd_batch **ptr, struct fd_batch *batch)
 233 {
 234         struct fd_batch *old_batch = *ptr;
 235         if (pipe_reference_described(&(*ptr)->reference, &batch->reference,
 236                         (debug_reference_descriptor)__fd_batch_describe))
 237                 __fd_batch_destroy(old_batch);
 238         *ptr = batch;
 239 }
 240
 241 /* fwd-decl prototypes to untangle header dependency :-/ */
 242 static inline void fd_context_assert_locked(struct fd_context *ctx);
 243 static inline void fd_context_lock(struct fd_context *ctx);
 244 static inline void fd_context_unlock(struct fd_context *ctx);
 245
 246 static inline void
 247 fd_batch_reference_locked(struct fd_batch **ptr, struct fd_batch *batch)
 248 {
 249         struct fd_batch *old_batch = *ptr;
 250
 251         if (old_batch)
 252                 fd_context_assert_locked(old_batch->ctx);
 253         else if (batch)
 254                 fd_context_assert_locked(batch->ctx);
 255
 256         if (pipe_reference_described(&(*ptr)->reference, &batch->reference,
 257                         (debug_reference_descriptor)__fd_batch_describe)) {
 258                 struct fd_context *ctx = old_batch->ctx;
 259                 fd_context_unlock(ctx);
 260                 __fd_batch_destroy(old_batch);
 261                 fd_context_lock(ctx);
 262         }
 263         *ptr = batch;
 264 }
 265
 266 #include "freedreno_context.h"
 267
 268 static inline void
 269 fd_reset_wfi(struct fd_batch *batch)
 270 {
 271         batch->needs_wfi = true;
 272 }
 273
 274 void fd_wfi(struct fd_batch *batch, struct fd_ringbuffer *ring);
 275
 276 /* emit a CP_EVENT_WRITE:
 277  */
 278 static inline void
 279 fd_event_write(struct fd_batch *batch, struct fd_ringbuffer *ring,
 280                 enum vgt_event_type evt)
 281 {
 282         OUT_PKT3(ring, CP_EVENT_WRITE, 1);
 283         OUT_RING(ring, evt);
 284         fd_reset_wfi(batch);
 285 }
 286
 287 #endif /* FREEDRENO_BATCH_H_ */